Human Motion Prediction Using Spatial Semantics of Objects: The PADSO Approach

Human Motion Prediction (HMP) is a key capability for a wide range of applications, including collaborative robotics, rehabilitation, and human-machine interaction. Despite advances in deep learning for HMP, most existing methods overlook the environmental context. In this paper, we present Prediction of Actions through Data about a Single Object (PADSO), a novel HMP model that enriches input human poses with spatial semantic information by incorporating environmental object data, thereby enabling joint processing of human pose and object pose predictions. Experimental evaluation on the GRasping Actions with Bodies (GRAB) dataset shows that PADSO significantly improves prediction accuracy over both a non-spatially aware version of the same architecture and the Zero-Velocity baseline. These results highlight the critical role of spatial context in enhancing the robustness of HMP. Furthermore, the efficient design of PADSO enables real-time inference, a critical feature for deployment in real-world scenarios.